Access Trends of In-network Cache for Scientific Data

Ruize Han, Alex Sim, Kesheng Wu, Inder Monga, Chin Guok, Frank Würthwein, Diego Davila, Justas Balcas, Harvey B. NewmanSNTA@HPDC 2022

Abstract

As scientific experiments like those at the Large Hadron Collider (LHC) reach exabyte scales, the distribution of data across wide-area networks (WAN) becomes a critical bottleneck. This paper evaluates the effectiveness of the Southern California Petabyte Scale Cache (SoCal Repo), a federated storage system designed to mitigate network congestion. Through a longitudinal study of 7.5 million data access requests, we analyze the impact of regional caching on traffic reduction and the predictability of access patterns using Long Short-Term Memory (LSTM) networks. Our findings demonstrate that while regional caches can reduce WAN traffic by up to 57%, performance is highly sensitive to shifts in user behavior. Furthermore, we show that cache utilization is predictable with 88.4% accuracy, suggesting a path toward automated, dynamic resource allocation in scientific data networks.

Introduction

Scientific discovery in high-energy physics (HEP) is increasingly dependent on the efficient movement of massive datasets. Traditional tiered storage models often struggle with the volume of requests, leading to congestion on backbone infrastructures like ESnet. To address this, "Data Lakes"—distributed regional caches—are deployed closer to end-users to serve as a buffer. This study investigates the SoCal Repo to quantify the actual network load reduction provided by such caches and to determine if access patterns are sufficiently predictable to support proactive network management.

Methodology

The researchers conducted a longitudinal analysis of the SoCal Repo, which integrates 24 cache nodes across Caltech, UCSD, and ESnet.

System Architecture

The infrastructure utilizes XCache, based on the XRootD protocol. The system operates by intercepting user file requests; a local redirector determines if the data exists within the regional cache (a cache hit). If the data is absent (a cache miss), it is fetched from a national repository and stored locally for future use.

Predictive Modeling

To forecast future demand, the study employed Long Short-Term Memory (LSTM) networks. The modeling process involved:

  • Feature Engineering: Tracking eight daily metrics, including access counts and data reuse rates.
  • Normalization: Applying z-score normalization to manage the high variance in data scales.
  • Smoothing: Implementing a Moving Average (MA) LSTM model to mitigate the impact of extreme outliers in data transfer sizes, which often skew standard machine learning performance.

Dataset and Analysis

The study analyzed monitoring data collected between July 2021 and January 2022, encompassing approximately 7.5 million requests. The data was split into an 80% training set and a 20% testing set.

Access Pattern Insights

Analysis revealed several critical trends in how researchers interact with scientific data:

  1. Seasonality: Periodogram analysis identified strong cycles at 7, 31, and 62 days, corresponding to weekly work patterns and monthly project milestones.
  2. The "Streaming" Shift: A significant drop in cache effectiveness (from ~60% to ~6% traffic reduction) occurred toward the end of the study. This was traced to a shift in user behavior from reusing existing datasets to "streaming" new data files, highlighting the cache's sensitivity to different research phases.
  3. Data Reuse vs. Hit Rate: While the overall hit rate fluctuated with streaming behavior, the reuse of specific popular files remained relatively stable, suggesting that a small subset of data often drives the majority of cache value.

Results

The SoCal Repo demonstrated a significant capacity for network optimization:

  • Traffic Reduction: Under optimal conditions, the system reduced WAN traffic by 57%. Even when accounting for unpredictable shifts in user behavior, the average reduction remained at 23%.
  • Prediction Accuracy: The LSTM models achieved an overall accuracy of 88.4% in predicting daily cache utilization. Notably, the models were more effective at predicting the number of accesses (over 90% accuracy) than the volume of data transferred, as the latter is more susceptible to bursty, unpredictable spikes.
  • Bandwidth Efficiency: Prior to the shift toward streaming, the network traffic demand reduction rate reached 2.35, effectively doubling the utility of the available backbone bandwidth.

Conclusion and Future Work

This research confirms that regional caches are a viable solution for reducing the burden on scientific network backbones. While LSTMs can accurately predict general trends and human-centric cycles, they still struggle with "black swan" events—atypical, massive data transfers.

Future efforts should focus on extending the dataset duration to capture multi-year scientific cycles and testing these models across other disciplines like genomics. The ultimate goal is the implementation of dynamic, "in-network" management systems that can automatically reconfigure storage and bandwidth based on predicted demand, moving beyond passive monitoring to active network optimization.

Summary updated

BibTeX
@inproceedings{han2022access,
  author = {Ruize Han and Alex Sim and Kesheng Wu and Inder Monga and Chin Guok and Frank Würthwein and Diego Davila and Justas Balcas and Harvey B. Newman},
  title = {{Access Trends of In-network Cache for Scientific Data}},
  booktitle = {SNTA@HPDC},
  year = {2022},
}
Figure